refactor(fa2): expose a Python-free C runtime

[LiangSu8899]

Summary

• move the existing vendored FA2 C entry points into a Python-free shared library
• keep flash_rt_fa2 as a thin Python adapter over the same implementation
• install the adapter and raw library as one relocatable package unit
• centralize the existing C declarations in one shared header

Design boundaries

• this is shared attention infrastructure, not model-specific runtime behavior
• existing fvk_attention_fa2_fwd_* signatures and the compiled kernel matrix remain unchanged
• the raw library is an internal native linkage boundary, not a new public versioned ABI
• native and Python producers consume one implementation and one symbol set

Packaging

flash_rt_fa2 and libflashrt_fa2_raw.so must be installed in the same directory. Both use an $ORIGIN runtime search path. Deployments that copied only the Python extension must switch to the complete install unit.

Validation

• isolated SM120 build of the raw library and Python adapter
• expected C symbols exported with no unresolved Python or FA2 entry symbols
• adapter dynamic dependency and $ORIGIN runtime paths verified
• BF16 attention executed through the Python adapter on SM120
• install-unit and import-without-upstream-FA2 smoke tests
• diff contains no model semantics, private paths, environment details, or generated artifacts

Merge order

Merge this shared-infrastructure PR before #144. Rebase #144 afterward so the attention-owner change is reviewed and landed exactly once.